[core] Fix a memory leak due to lineage counting (#31488)

Assume two objects A and B and B depends on A. If A got delete first, it's won't be released because we need to preserve it for B.

The bug here is that even B is released, A won't be released and thus memory leak.

This PR fixed this issue by force delete the object when the ref goes to 0.

python/ray/tests/test_reference_counting_2.py

@@ -777,6 +777,27 @@ def remote_generator():
         _fill_object_store_and_get(r_oid, succeed=False)
 
 
+def test_lineage_leak(shutdown_only):
+    ray.init()
+
+    @ray.remote
+    def process(data):
+        return b"\0" * 100_000_000
+
+    data = ray.put(b"\0" * 100_000_000)
+    ref = process.remote(data)
+    ray.get(ref)
+    del data
+    del ref
+
+    def check_usage():
+        from ray._private.internal_api import memory_summary
+
+        return "Plasma memory usage 0 MiB" in memory_summary(stats_only=True)
+
+    wait_for_condition(check_usage)
+
+
 if __name__ == "__main__":
     import sys
 

src/ray/core_worker/reference_count.cc

@@ -478,10 +478,9 @@ int64_t ReferenceCounter::ReleaseLineageReferences(ReferenceTable::iterator ref)
     RAY_LOG(DEBUG) << "Releasing lineage internal for argument " << argument_id;
     arg_it->second.lineage_ref_count--;
     if (arg_it->second.ShouldDelete(lineage_pinning_enabled_)) {
-      // We only decremented the lineage ref count, so the argument value
-      // should already be released.
       RAY_CHECK(arg_it->second.on_ref_removed == nullptr);
       lineage_bytes_evicted += ReleaseLineageReferences(arg_it);
+      ReleasePlasmaObject(arg_it);
       EraseReference(arg_it);
     }
   }